API Reference: utils

grilly.utils provides data loading, tensor conversion, checkpointing, device management, and HuggingFace integration.

Data Loading

Class	Signature	Description
`Dataset`	`Dataset()`	Abstract base class. Override `__len__` and `__getitem__`.
`TensorDataset`	`TensorDataset(*tensors)`	Dataset from numpy arrays.
`ArrayDataset`	`ArrayDataset(*arrays)`	Dataset from arrays (alias).
`Subset`	`Subset(dataset, indices)`	Subset of a dataset by indices.
`ConcatDataset`	`ConcatDataset(datasets)`	Concatenation of datasets.
`DataLoader`	`DataLoader(dataset, batch_size=1, shuffle=False, num_workers=0)`	Iterable batch loader.
`random_split`	`random_split(dataset, lengths)`	Split dataset into random subsets.

Class	Description
`RandomSampler`	Sample indices randomly.
`SequentialSampler`	Sample indices in order.
`BatchSampler`	Wrap a sampler to yield batches.

Class	Description
`Compose`	Chain transforms.
`ToFloat32`	Convert to float32.
`Normalize`	Normalize with mean and std.
`Flatten`	Flatten to 1D.
`RandomNoise`	Add random noise (augmentation).
`RandomFlip`	Random horizontal flip (augmentation).
`OneHot`	One-hot encode labels.
`Lambda`	Apply an arbitrary function.

Function	Description
`to_vulkan(array)`	Upload numpy array to GPU as VulkanTensor.
`to_vulkan_batch(arrays)`	Batch upload multiple arrays.
`from_vulkan(vt)`	Download VulkanTensor to numpy.
`ensure_vulkan_compatible(array)`	Ensure array is float32 and contiguous.

Class	Description
`VulkanTensor`	Wrapper around a GPU buffer. Supports zero-copy access.

Function	Signature	Description
`save_checkpoint`	`save_checkpoint(model, optimizer, epoch, path)`	Save model + optimizer state.
`load_checkpoint`	`load_checkpoint(model, optimizer, path)`	Load model + optimizer state.

Function	Description
`get_device()`	Get the current device name.
`set_device(index)`	Set the active GPU by index.
`device_count()`	Number of available Vulkan devices.

Class	Description
`DeviceManager`	Multi-backend device manager (Vulkan + CUDA).

Function	Description
`xavier_uniform_(array)`	Xavier/Glorot uniform initialization.
`xavier_normal_(array)`	Xavier/Glorot normal initialization.
`kaiming_uniform_(array)`	Kaiming/He uniform initialization.
`kaiming_normal_(array)`	Kaiming/He normal initialization.

Class	Description
`HuggingFaceBridge`	Load pretrained HuggingFace weights without PyTorch runtime.

from grilly.utils import HuggingFaceBridge

bridge = HuggingFaceBridge()
weights = bridge.load_weights("bert-base-uncased")

Class	Description
`OnnxModelLoader`	Load ONNX models into grilly modules.
`OnnxExporter`	Export grilly models to ONNX format.
`OnnxFineTuner`	Fine-tune ONNX models with grilly optimizers.
`GrillyOnnxModel`	Grilly wrapper for loaded ONNX models.

Symbol	Description
`Tensor`	Drop-in Tensor class with PyTorch-compatible API.
`tensor`, `zeros`, `ones`, `randn`	Factory functions matching PyTorch signatures.

Class	Description
`StreamingPipeline`	Batched embed + upload pipeline for inference.

Requires matplotlib.

Function	Description
`plot_training_history(history)`	Plot loss/accuracy curves.
`plot_gradient_flow(model)`	Visualize gradient magnitudes per layer.
`print_model_summary(model)`	Print model architecture summary.